March, 2013

Agric Eng Int: CIGR Journal

Open access at http://www.cigrjournal.org

Vol. 15, No.1 211

Detection of unhealthy region of plant leaves and classification of plant leaf diseases using texture features S. Arivazhagan, R. Newlin Shebiah*, S. Ananthi, S. Vishnu Varthini (Department of Electronics and Communication Engineering, Mepco Schlenk Engineering College, Sivakasi Tamilnadu, 626 005, India) Abstract: Plant diseases have turned into a dilemma as it can cause significant reduction in both quality and quantity of agricultural products. Automatic detection of plant diseases is an essential research topic as it may prove benefits in monitoring large fields of crops, and thus automatically detect the symptoms of diseases as soon as they appear on plant leaves. The proposed system is a software solution for automatic detection and classification of plant leaf diseases.

The developed

processing scheme consists of four main steps, first a color transformation structure for the input RGB image is created, then the green pixels are masked and removed using specific threshold value followed by segmentation process, the texture statistics are computed for the useful segments, finally the extracted features are passed through the classifier. The proposed algorithm’s efficiency can successfully detect and classify the examined diseases with an accuracy of 94%. Experimental results on a database of about 500 plant leaves confirm the robustness of the proposed approach. Keywords: HSI, color co-occurrence matrix, texture, SVM, plant leaf diseases Citation: S.Arivazhagan, R. Newlin Shebiah, S.Ananthi, S.Vishnu Varthini. 2013. leaves and classification of plant leaf diseases using texture features.

Detection of unhealthy region of plant

Agric Eng Int: CIGR Journal, 15(1): 211－217.

Introduction

plant diseases were devastating, some of the crop

Images form important data and information in

plant disease losses in Georgia (USA) is approximately

biological sciences. Digital image processing and image

$653.06 million (Jean, 2009). In India no estimation has

analysis

in

been made but it is more than USA because the preventive

microelectronics and computers has many applications in

steps taken to protect our crops are not even one-tenth of

biology and it circumvents the problems that are

that in USA.

1

technology

cultivation has been abandoned. It is estimated that 2007

based

on

the

advances

associated with traditional photography. This new tool

The naked eye observation of experts is the main

helps to improve the images from microscopic to

approach

telescopic range and also offers a scope for their analysis.

identification of plant diseases.

It, therefore, has many applications in biology (Rastogi

continuous monitoring of experts which might be

and Chadda, 1989).

prohibitively expensive in large farms. Further, in some

.

developing countries, farmers may have to go long

Plant diseases cause periodic outbreak of diseases

which leads to large scale death and famine.

adopted

in

practice

for

detection

and

But, this requires

It is

distances to contact experts, this makes consulting experts

estimated that the outbreak of helminthosporiose of rice in

too expensive and time consuming (Al-Hiary et al., 2011)

north eastern India in 1943 caused a heavy loss of food

and moreover farmers are unaware of non-native diseases.

grains and death of a million people. Since the effects of

Automatic detection of plant diseases is an important research topic as it may prove benefits in monitoring large

Received date: 2012-11-23 Accepted date: 2013-02-21 * Corresponding author: R. Newlin Shebiah, Email: [email protected].

fields of crops, and thus automatically detect the diseases from the symptoms that appear on the plant leaves. This enables machine vision that is to provide image based

212

March

Agric Eng Int: CIGR Journal

Open access at http://www.cigrjournal.org

Vol. 15, No.1

automatic inspection, process control and robot guidance.

may be broadly classified into three types.

Comparatively, visual identification is labor intensive, less

bacterial, fungal and viral diseases.

accurate and can be done only in small areas. Kim et al. (2009) have classified the grape fruit peel diseases using color texture features analysis.

2

They are

Proposed methodology First, the images of various leaves are acquired using a

The

texture features are calculated from the Spatial Gray-level

digital camera.

Dependence Matrices (SGDM) and the classification is

applied to the acquired images to extract useful features

done using squared distance technique. Grape fruit peel

that are necessary for further analysis. After that, several

might be infected by several diseases like canker, copper

analytical techniques are used to classify the images

burn, greasy spot, melanose and wind scar (Kim et al.,

according to the specific problem at hand.

2009).

depicts the basic procedure of the proposed vision-based

Helly et al. (2003) developed a new method in which

Then image-processing techniques are

Figure 1

detection algorithm in this paper.

Hue Saturation Intensity (HIS) - transformation is applied to the input image, then it is segmented using Fuzzy C-mean algorithm. Feature extraction stage deals with the color, size and shape of the spot and finally classification is done using neural networks (Helly et al.,

Figure 1 Block diagram of proposed approach

2003). Real time specific weed discrimination technique using multilevel wavelet decomposition was proposed by Siddiqil et al. (2009). In this histogram equalization is

In the initial step, the RGB images of all the leaf samples were picked up.

used for preprocessing. Features are extracted from

The step-by-step procedure of the proposed system:

wavelet decomposition and finally classified by Euclidean

1) RGB image acquisition;

distance method (Siddiqil et.al, 2009)

2) Convert the input image from RGB to HSI format;

Al-Bashish et al. (2011) developed a fast and accurate

3) Masking the green-pixels;

method in which the leaf diseases are detected and

4) Removal of masked green pixels;

classified using k-means based segmentation and neural

5) Segment the components;

networks based classification.

6) Obtain the useful segments;

Automatic classification

of leaf diseases is done based on high resolution multispectral and stereo images (Bauer et al., 2011).

7) Computing the texture features using Color-CoOccurrence methodology;

Sugar beet leaves are used in this approach. Segmentation is the process that is carried out to extract the diseased region and the plant diseases are

8) Configuring the Neural Networks for Recognition. 2.1

Color transformation structure First, the RGB images of leaves are converted into HSI

graded by calculating the quotient of disease spot and leaf

color space representation.

areas. An optimal threshold value for segmentation can

space is to facilitate the specification of colors in some

be obtained using weighted Parzen-window (Jun and

standard, generally accepted way. HSI (hue, saturation,

Wang, 2008).

This reduces the computational burden

intensity) color model is a popular color model because it

and storage requirements without degrading the final

is based on human perception (Gonzalez and Woods,

segmentation results.

2008). Hue is a color attribute that refers to the dominant

The purpose of the color

In this paper, detection and classification of leaf

color as perceived by an observer. Saturation refers to

diseases has been proposed, this method is based on

the relative purity or the amount of white light added to

masking and removing of green pixels, applying a specific

hue and intensity refers to the amplitude of the light.

threshold to extract the infected region and computing the

Color spaces can be converted from one space to another

texture statistics to evaluate the diseases. Plant diseases

easily.

After the transformation process, the H

March, 2013 Detection of unhealthy region of plant leaves and classification of plant leaf diseases using texture features Vol. 15, No.1 213

component is taken into account for further analysis. S

information.

and I components are dropped since it does not give extra

components.

a. Input image infected by bacterial brown spot

b. Hue component

Figure 2 shows the H, S and I

c. Saturation component

d. Intensity component

Figure 2 HSV components of a image infected by brown spots

2.1.1

gray-levels occur in relation to other gray levels (Argenti

Masking green pixels

In this step we identify the mostly green colored

et al., 2008).

These matrices measure the probability

After that, based on specified threshold value

that a pixel at one particular gray level will occur at a

that is computed for these pixels, the mostly green pixels

distinct distance and orientation from any pixel given that

are masked as follows: if the green component of the

pixel has a second particular gray level.

pixel intensity is less than the pre-computed threshold

are represented by the function P(i, j, d, θ) where i

value, the red, green and blue components of the this

represent the gray level of the location (x, y), and j

pixel is assigned to a value of zero.

represents the gray level of the pixel at a distance d from

pixels.

This is done in

sense that the green colored pixels mostly represent the

location (x, y) at an orientation angle of θ.

healthy areas of the leaf and they do not add any valuable

generated for H image.

weight to disease identification.

2.4

Furthermore this

In this step, the pixels with zeros red, green, blue gives

more

accurate

disease

This is helpful as it classification

and

significantly reduces the processing time. 2.2

From the above steps, the infected portion of the leaf The infected region is then segmented into

a number of patches of equal size.

The size of the patch

is chosen in such a way that the significant information is not lost. taken. Not

computed for the H image as given in Equations (1) to (5).

Contrast   i , j  0 (i, j ) 2 C (i, j )

(1)

Energy   i , j  0 C (i, j ) 2

(2)

N 1

N 1

Segmentation:

is extracted.

Texture features

homogeneity, Cluster shade and cluster prominence are

Removing the masked cells

values were completely removed.

In this approach patch size of 32×32pixels is

Local Homogeneity   i , j  0 C (i, j ) / (1  (i  j )2 )

(3)

Cluster Shade   i , j  0 (i  M x  j  M y )3 C (i, j )

(4)

N 1

N 1

Cluster Prominence   i , j  0 (i  M x  j  M y )4 C (i, j ) N 1

(5)

The next step is to extract the useful segments. all

segments

contain

significant

amount

From the texture features, the plant diseases are

of

information. So the patches which are having more than

classified into various types.

fifty percent of the information are taken into account for

2.5

the further analysis.

2.5.1 Minimum distance criterion

2.3

SGDM’s are

Texture features like Contrast, Energy, Local

significantly reduces the processing time. 2.1.2

The SGDM’s

Classifier

In the classification phase, the co-occurrence features

Color co-occurrence method The color co-occurrence texture analysis method is

for the leaves are extracted and compared with the

The gray level

corresponding feature values stored in the feature library.

co-occurrence methodology is a statistical way to

The classification is first done using the Minimum

describe shape by statistically sampling the way certain

Distance Criterion - (Arivazhagan et al., 2010).

developed through the SGDM.

The

214

March

Agric Eng Int: CIGR Journal

Open access at http://www.cigrjournal.org

Vol. 15, No.1

success of classification is measured using the classification

Classification of new instances for the one-versus-all case

gain (G) and is calculated using Equation (6).

is done by a winner-takes-all strategy, in which the

G (%) 

Ccorr  100 M

(6)

where, Ccorr is the number of images correctly classified and M is the total number of images belonging to the

class (Chamasemani and Singh, 2011).

3

Results and discussion About 500 plant leaves of 30 different native plant

particular texture group. 2.5.2

classifier with the highest output function assigns the

species of Tamil Nadu have been collected for our

SVM classifier

Support vector machines (SVMs) are a set of related

approach.

The acquired leaf images are converted into

supervised learning methods used for classification and

HSI format.

regression.

Supervised learning involves analyzing a

energy, local homogeneity, shade and prominence are

given set of labeled observations (the training set) so as to

derived from the co-occurrence matrix. With these set

predict the labels of unlabelled future data (the test set).

of co-occurrence features the plant diseases are detected.

Specifically, the goal is to learn some function that

Samples of leaves with various diseases like early scorch,

describes the relationship between observations and their

yellow spots, brown spots, late scorch, bacterial and

labels (Chi & Lin, 2002).

fungal diseases are shown in Figure 3.

More formally, a support

The co-occurrence features like contrast,

vector machine constructs a hyper plane or set of hyper planes in a high- or infinite-dimensional space, which can be used for classification, regression, or other tasks. Intuitively, a good separation is achieved by the hyper plane that has the largest distance to the nearest training data point of any class (so-called functional margin), in

a. Bacterial disease in rose and beans leaf

general the larger the functional margin the lower the generalization error of the classifier. In the case of support vector machines, a data point is viewed as a p-dimensional vector (a list of p numbers), and we want to know whether we can separate such points with a (p − 1)-dimensional hyper plane.

This is

b. Sun burn disease in lemon leaf

c. Early scorch disease in banana leaf

d. Late scorch disease in beans leaf

e. Fungal disease in beans leaf

called a linear classifier. There are many hyper planes that might classify the data.

One reasonable choice as

the best hyper plane is the one that represents the largest separation, or margin, between the two classes.

So we

choose the hyper plane so that the distance from it to the nearest data point on each side is maximized.

Figure 3

Sample images of infected leaves

Multiclass SVM aims to assign labels to instances by using support vector machines, where the labels are drawn from a finite set of several elements.

The

As a sample, a rose leaf that is infected by bacterial disease is given as input to the algorithm.

Color

dominant approach for doing so is to reduce the single

transformation structure on the input image is performed.

multiclass problem into multiple binary classification

Then the green pixels are masked and removed using a

problems.

Common methods for such reduction include:

specific threshold value. Then the R, G, B components

building binary classifiers which distinguish between (i)

are mapped to the thresholded image. These steps are

one of the labels and the rest (one-versus-all) or (ii)

shown in Figure 4.

between

affected by various diseases.

every

pair

of

classes

(one-versus-one).

Table 1 lists the set of leaves that are

March, 2013 Detection of unhealthy region of plant leaves and classification of plant leaf diseases using texture features Vol. 15, No.1 215

a. Input image

b. Hue content

Figure 4

c. Threshold image

d. R component mapped output

Detection of infected region for a rose leaf

Table 1 Detected diseased region of various leaves Plant species

Input image

Hue content

Thresholded image

R component mapped output

Beans

Lemon

Banana

Guava

After mapping the R, G, B components of the input image to the thresholded image, the co-occurrence features are calculated.

used for training, testing and classification gain for each type of leaves is shown in Table 2.

The co-occurrence features for

The classification gain obtained by Minimum

the leaves are extracted and compared with the

Distance Criterion is 86.77%. The detection accuracy is

corresponding feature values stored in the feature library.

improved to 94.74% by SVM classifier.

The classification is first done using the Minimum

and the testing sets for each type of leaf along with their

Distance Criterion.

detection accuracy is shown in Table 2.

The leaf images are divided into

training and testing set, where 5% of the leaf images from each group are used to train the system and the remaining images serves as the testing set.

The number of images

The training

From the results it can be seen that the detection accuracy is enhanced with SVM classifier. The two class problem is then extended to multiclass

216

March

Agric Eng Int: CIGR Journal

Open access at http://www.cigrjournal.org

Vol. 15, No.1

problem where the detected leaf diseases are then

testing sets for each type of leaf disease along with their

classified into various categories.

detection accuracy is shown in Table 3.

Table 2

Training and the

Comparison of results by minimum distance classifier and support vector machine Detection accuracy/%

No. of images used for training

No. of images used for testing

Banana

10

10

82.25

90

Beans

10

12

96.43

91.66 92.86

Plant species

MDC

SVM

Guava

10

14

78.95

Jackfruit

10

10

82.35

100

Lemon

10

30

94.4

96.66

Mango

10

17

80

94.12

Potato

10

10

78.57

100

Tomato

10

27

95.24

92.59

86.77

94.74

Overall accuracy

Table 3 Plant Species Banana

Beans

Guava

Jackfruit

Lemon

Mango

Potato

Sapota

Tomato

Results of leaf disease recognition system

Category

No. of Images used for Training

No. of Images used for Testing

Good

5

6

Late scorch

5

7

Good

4

9

Bacterial spot

2

4

Fungal spot

4

11

Good

5

7

Chocolate spot

5

7

Good

4

5

Bacterial disease

4

3

Fungal disease

2

2

Good

4

5

Bacterial disease

4

20

Sun burn

2

3 6

Good

3

Bacterial disease

3

4

Sooty mold

4

8

Good

4

7

Early blight

3

9

Late blight

3

12

Good

4

5

Scorch

3

3

Ashen mold

3

2

Good

4

5

Bacterial disease

4

19

Leaf lesion

2

4

Detection Accuracy 84.60%

87.50%

92.86%

90%

82.14%

83.33%

96.43%

80%

82.15%

Overall Accuracy

4 Conclusion

87.66%

lemon, mango, potato, tomato, and sapota.

The diseases

specific to those plants were taken for our approach.

An application of texture analysis in detecting and

The experimental results indicate the proposed approach

classifying the plant leaf diseases has been explained in

can recognize and classify the leaf diseases with a little

this paper. Thus the proposed algorithm was tested on

computational effort.

ten species of plants namely banana, beans, jackfruit,

can be identified at the initial stage itself and the pest

By this method, the plant diseases

March, 2013 Detection of unhealthy region of plant leaves and classification of plant leaf diseases using texture features Vol. 15, No.1 217

control tools can be used to solve pest problems while

deletion), also the taken feature identification vectors

minimizing risks to people and the environment.

need to further optimized.

The

In order to improve disease

reasons for misclassification are as follows: the

identification rate at various stages, the training samples

symptoms of the diseased plant leaves vary (at the

can be increased and shape feature and color feature

beginning, tiny, dark brown to black spots, at later time, it

along with the optimal features can be given as input

has the phenomena of withered leaf, black or part leaf

condition of disease identification.

References Al-Bashish, D., M. Braik, and S. Bani-Ahmad. 2011. Detection and classification of leaf diseases using K-means-based segmentation and neural networks based classification. Information Technology Journal, 10(2): 267-275. 2011.

Fast and accurate detection and

classification of plant diseases.

International Journal of

Computer Applications, 17(1): 31-38. Fast algorithms

for texture analysis using co-occurrence matrices.

IEEE

proceedings, 137, (6): 443-448. 2010.

Classification of citrus and non-citrus fruits Computing Communication and

Networking Technologies, ICCCNT-2010. Bauer, S. D., F. Korc, W. Forstner.

An

and diagnosis. in Proc. IICAI,

pp.1182-1195.

Jean, W. 2009. Extension plant pathologist, georgia plant disease loss estimates, www.caes.uga.edu/publications. parzen window estimation.

Journal of Applied Sciences,

8(5):772-779. Kim, D. G., T. F. Burks, J. Qin, and D. M. Bulanon.

Arivazhagan, S., R. N. Shebiah, S. S. Nidhyanandhan, and L. using texture features.

Neural Networks, 13(2): 415-425. Helly, M. E., A. Rafea, and Salwa-El-Gammal. 2003.

Jun, W., and S. Wang. 2008. Image thresholding using weighted

Argenti, F., L. Alparone, and G. Benelli. 1990.

Ganesan.

IEEE Transactions on

integrated image processing system for leaf disease detection

Al-Hiary, H., S. Bani-Ahmad, M. Reyalat, M. Braik, and Z. AlRahamneh.

multiclass support vector machines.

2011.

Classification of grapefruit peel diseases using color texture feature analysis.

International Journal on Agriculture and

Biological Engineering, 2(3): 41-50. Rastogi, R., and V. K. Chadda.

The potential of

automatic methods of classification to identify leaf diseases from multispectral images. Precision Agriculture, 12: 361-377.

2009.

Applications of image

processing in biology and agriculture J. K. Sainis, Molecular Biology and Agriculture Division, BARC newsletter. Siddiqil, M. H., S. Sulaiman, I. Faye, and I. Ahmad.

2009. A

Chamasemani, F. F., and Y. P. Singh. Malaysia, Bio-Inspired

real time specific weed discrimination system using multi-level

Computing: Theories and Applications (BIC-TA), 2011 Sixth

wavelet decomposition. International Journal of Agriculture

International Conference.

Faculty of Information Technology,

MultiMedia University. Cyberjaya, Malaysia. Chih-Wei, H., and C. Lin.

2002. A comparison of methods for

and Biology, 11(5): 559-565. Gonzalez, R., R. E. Woods. 2008 Digital image processing. Third edition, Pearson Education, Prentice-Hall, Inc.